The Humanitarian Imperative for Minimally-Just AI in Weapons
Jason Scholz, Jai Christian Galliott · 2021
Abstract For the use of force to be lawful and morally just, future autonomous systems must not commit humanitarian errors or acts of fratricide. To achieve this, we distinguish a novel preventative form of minimally-just autonomy using artificial intelligence (MinAI) to avert attacks on protected symbols, protected sites, and signals of surrender. MinAI compares favorably with respect to maximally-just forms proposed to date. We examine how fears of speculative artificial general intelligence has distracted resources from making current weapons more compliant with international humanitarian law, particularly Additional Protocol 1 of the Geneva Convention and its Article 36. Critics of our approach may argue that machine learning can be fooled, that combatants can commit perfidy to protect themselves, and so on. We confront this issue, including recent research on the subversion of AI, and conclude that the moral imperative for MinAI in weapons remains undiminished.